{"id":"ab03bdb3-2e6e-4d19-ba26-3e1652856cfe","arxiv_id":"2412.09978","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A congestion-aware, price-aware charging scheduler improves simulated electric ride-hailing profit by up to about 15% and customer service rate by up to about 19 percentage points versus threshold-based benchmark policies.","lead":"A new charging scheduler for electric ride-hailing fleets plans when and where vehicles should charge to avoid queueing at stations and to use cheaper off-peak electricity. In simulated tests with 100 vehicles, it increased daily profit and the share of customers served relative to four benchmark charging policies.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"P1's day-ahead plan uses waiting times and profit rates estimated from the Fastest benchmark, not from CongestionAware itself; without a fixed-point check, the reported profit/service gains may not reflect the policy's own equilibrium.","rationale":"Good-faith reading: the paper's core claim is that a sequential MILP day-ahead plan plus online reactive assignment raises profit and service rate relative to four benchmark policies. The simulation design is largely coherent: the comparison is on identical demand instances, standard deviations (Table 5) are small relative to the reported gaps, and the paper includes extensive sensitivity analysis. I do not find a fatal flaw in the optimization models themselves. The weakest link is the parameterization of P1: the waiting-time signal Wbar_hs and the opportunity-cost parameter gamma come from the Fastest benchmark, not from the policy being planned. Because the whole method is built around being congestion-aware, using a congestion signal from a different queuing regime is an arbitrary choice. The online P3 uses realized queuing information, which mitigates but does not eliminate the mismatch: P1 determines which vehicles enter the charging pool and their target SoCs, so a miscalibrated P1 can shift the operating point of the entire policy. The paper reports no fixed-point or consistency check, so the quantitative headline gains are conditional on this initialization. The concrete test above would settle whether the gains persist under self-consistent estimates. I also flag the abstract's unreproducible numbers ('OptChg' does not exist in the full text) as a supporting reason to keep the verdict conditional, though it is not the primary technical concern. The reader's weakest assumption is essentially the same issue, and the proposed test would resolve it; the verdict should remain CONDITIONAL until such a check is provided.","tokens_in":38712,"tokens_out":13414,"duration_ms":161226,"concrete_test":"Implement a fixed-point iteration: (i) run Fastest on the 10 training days to initialize Wbar_hs and gamma; (ii) solve P1 and simulate CongestionAware on those training days; (iii) recompute Wbar_hs and gamma from the realized CA charging sessions and profits; (iv) re-solve P1 and re-simulate on the 5 test days; repeat until the parameters change by less than 1%. Report the resulting KPIs from Table 4. If profit or service rate relative to DynaThreshold changes by more than about 1 percentage point, or the ordering reverses, the reported gains depend on the unvalidated initialization. A cheaper auxiliary check is to rerun the comparison with Wbar_hs set to zero to bound the effect of the waiting-time estimate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2.a states that the day-ahead planning parameters Wbar_hs and gamma are obtained by simulating the Fastest charging policy on 10 training days, and these are then used in P1's objective (Eq. (1)). Under CongestionAware, however, the realized charging waiting times are far smaller than under Fastest (Table 4: TW = 36.0/48.4 h for CA vs 77.1-136.1 h across the four benchmarks), so the cost coefficients in P1 are not the ones the policy actually induces. The paper provides no iterative re-estimation, no comparison with self-consistent Wbar, and no sensitivity analysis to the initial estimate. This matters because P1's schedule (which vehicles charge, when, and to what target SoC) is the distinguishing component of the method; if P1 were re-solved using waiting times realized under CA, the plan would change and the reported 7.65-15.05% profit and 7-12.3 percentage-point service-rate advantages could move. A secondary consistency issue is that the abstract's headline gains (+19.32% profit, +20.03% service rate; +3.91%/+4.60% vs 'OptChg') do not appear in Table 4 and 'OptChg' is never defined in the full text, so the quantitative claim is not fully reproducible from the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a sequential mixed-integer linear programming approach, named CongestionAware, for coordinating dispatching and charging of an electric ride-hailing fleet under stochastic demand, time-of-use energy prices, and congested charging stations. A day-ahead plan (Model P1) determines which vehicles charge, when, and to what target state of charge, using expected charger waiting times and average profit rates. During the day, an online vehicle-to-charger assignment (Model P3) and a batch dispatch model (Model P2) are solved every minute. The policy is tested in a Manhattan-like simulation with 100 vehicles, 12 chargers, and two demand scenarios (3000 and 4000 customers/day), and compared against four benchmark charging policies. The authors report that CongestionAware yields higher total profit and customer service rates, with lower charging waiting times, and support this with sensitivity analyses over battery capacity, charger counts, demand levels, and pricing schemes.","tokens_in":39082,"tokens_out":6424,"duration_ms":68057,"significance":"If the reported gains are robust, the paper offers a practically relevant tool: fleet operators could improve profit and service rates without adding charging infrastructure. The paper's strengths include explicit queueing modeling, minimum charging duration constraints, a day-ahead/online decomposition, comparison against four benchmark policies, and extensive sensitivity and scalability experiments. The main weaknesses are that the day-ahead plan is calibrated using waiting times and profit rates from a benchmark policy rather than from the proposed policy itself, the validation is based on only five days per scenario without significance testing, and the abstract's headline numbers are not reproduced in the body of the paper. These issues leave the central claim plausible but not yet established.","major_comments":[{"comment":"The day-ahead planning model P1 uses expected waiting times Wbar_hs and average profit rate gamma that are estimated by simulating the Fastest benchmark policy on 10 training days (Section 4.2.a). However, under the proposed CongestionAware policy the realized charging waiting times are much smaller than under Fastest (Table 4: TW = 36.0/48.4 h for CongestionAware versus 77.1–136.1 h for the benchmarks). Therefore the cost coefficients in Eq. (1) do not correspond to the congestion regime that the proposed policy actually induces. The paper provides no fixed-point or consistency check (re-estimating Wbar_hs and gamma under CongestionAware and re-solving P1), and no sensitivity analysis to these input estimates. Since P1's output is the distinguishing component of the method, the reported profit and service-rate advantages could shift if the plan were re-solved self-consistently. Please provide such a check or a sensitivity analysis over Wbar_hs and gamma.","section":"§4.2.a, Eq. (1)"},{"comment":"The quantitative claims in the abstract are not reproduced in the full text. The abstract states gains of up to +19.32% profit and +20.03% service rate, and relative to a benchmark called 'OptChg' gains of +3.91% profit and +4.60% service rate. Section 5 instead reports profit increases of 7.65%–10.69% (c3000) and 8.76%–15.05% (c4000), and service-rate increases of 7–10.8 and 7.9–12.3 percentage points (Table 4). The benchmark 'OptChg' is never defined in the full text, and no results for it appear in any table. These inconsistencies make the paper's main quantitative claims not reproducible from the manuscript; the authors should reconcile the abstract with the reported results and either define OptChg or remove it.","section":"Abstract vs. Section 5 / Table 4"},{"comment":"The stated replacement constraints for P3 in the case where the number of to-charge vehicles is smaller than the number of chargers require every charger to be assigned exactly one vehicle (sum_v x̂_vs = 1 for all s in Eq. (28)). This is infeasible whenever |Ω̃_t| < |S|, a case that occurs in the simulation. The intended constraints appear to be reversed or the implemented formulation differs from the text. Because P3 is the online assignment model used in every simulation step, this discrepancy is load-bearing for the reported results. Please clarify the correct formulation and confirm that the implemented model matches it.","section":"§3.3, Eqs. (27)–(28)"},{"comment":"All performance comparisons rest on five validation days per demand scenario, and no statistical significance tests are reported. While the standard deviations in Table 5 suggest that the largest differences are not noise, formal paired tests or confidence intervals are needed to support the claim of systematic outperformance, especially for the smaller differences among benchmark policies. This is particularly relevant because the validation sample is small and the demand process is stochastic. Please add significance tests or at least confidence intervals for the main KPIs (profit and service rate) in Tables 4 and 9–11.","section":"§4.2, Tables 4–5"}],"minor_comments":[{"comment":"The figure numbering is inconsistent: the timeline in Section 3.1 is called Figure 1, but the charging station map in Section 4.1 is also called Figure 1 while the displayed figure is numbered 'Figure 2. Charging station distribution in the study area.' Please re-number the figures in order of appearance.","section":"Figures 1–2"},{"comment":"The energy consumption rate is denoted μ in the models (Eqs. (15), (21)) but is listed as φ in Table 2; harmonize the notation.","section":"Table 2 and P2/P3 notation"},{"comment":"Charging power should be expressed in kW, not kWh: '50 kWh' should be '50 kW' and '11kWh' should be '11 kW' in the text and Table 2.","section":"§4.1, Table 2"},{"comment":"The paper reports that P1 is solved with optimality gaps of about 5%–14% within a one-hour time limit; please add an explicit caveat in the conclusions that the day-ahead plan is an approximate, not exact, solution, and discuss whether higher-quality solutions could change the results.","section":"§4.3.c and Appendix B"},{"comment":"The title on the first page of the manuscript ('Coordinated vehicle dispatching and charging scheduling for an electric ride-hailing fleet under charging congestion and dynamic prices') differs from the title under which the paper is listed ('Congestion-Aware Charging Coordination for Electric Ride-Hailing Fleets under Stochastic Demand'). Please align these before publication.","section":"Title"},{"comment":"In constraints (15) and (16), the set of requests is written as r_t in some places and R_t in others; the notation should be consistently R_t.","section":"§3.3, P2 constraints"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be a version whose abstract and title do not match the main text (the abstract refers to an undefined 'OptChg' benchmark and reports different gain magnitudes than Section 5). This suggests the abstract may have been carried over from another version; it must be reconciled before any decision. The calibration issue with Wbar_hs and gamma is the most substantive technical concern, since it affects the validity of the central comparison. If the authors can show that the advantages persist under a self-consistent estimation or a sensitivity analysis, the paper would be much stronger."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Plainly: this is a serious engineering paper with a genuine equilibrium-calibration problem. The authors build a three-layer MILP scheme — day-ahead charging plan (P1), reactive go-charge pool with end-of-day energy anticipation, and online vehicle-charger assignment (P3) — and they add realistic details that most prior work omits: minimum charging duration, maximum queue waiting with charger-chasing, partial recharge targets, ToU prices. The simulation description is detailed enough to be replicable in principle, and the online solve times are convincing.\n\nThe main claim, that CongestionAware beats four benchmarks in profit and service rate, is supported by Table 4: roughly 8-15% profit and 8-12 percentage-point service-rate gains. The sensitivity analysis is extensive — battery sizes, charger counts, demand from 1000 to 6000 customers/day, overnight charging costs, charger-type prices, plus a scaling study to 20,000 customers/day. That is real work, and the conclusion that the policy helps most when chargers are scarce is consistent with the mechanism.\n\nThe soft spot the stress-test note identifies is genuine. P1 uses expected waiting times Wbar_hs and average profit rate gamma estimated by simulating the Fastest benchmark on ten training days. Under CongestionAware, realized waiting times are much lower (36-48 hours vs 77-136 for benchmarks), so the cost coefficients in P1 are not the ones the policy induces. There is no fixed-point check, no sensitivity analysis over Wbar. If the planner were re-solved with self-consistent waiting times, the schedule would likely change. The direction of the bias actually works in the paper's favor — overestimating waiting costs makes the plan conservative about congestion — so the qualitative story probably holds. But the magnitude of the gains could move, and I would not treat the headline numbers as established.\n\nThere is also an abstract inconsistency that needs fixing. The top-of-paper abstract claims +19.32% profit and +20.03% service rate for 4000 customers, and +3.91%/+4.60% against something called 'OptChg'. The full-text abstract says up to +15.06% profit and +19.16% service, and Table 4 shows the strongest benchmark is DynaThreshold. 'OptChg' never appears in the paper. This is a version-control slip, but it makes the headline claim unreproducible from the tables.\n\nAlso worth noting: no code or data is shipped, and the validation uses five days without significance tests. The standard deviations in Table 5 are small, so the differences are probably real, but a test would be cheap to add.\n\nRecommendation: send to review. The contribution is new, the modeling is detailed, and the issues are addressable — a self-consistency check on Wbar, a significance test, and a corrected abstract. I'd want those before the numbers are cited, but the paper deserves referee time.","headline":"A detailed simulation study of a congestion-aware charging policy whose headline gains are plausible but rest on a day-ahead planner calibrated to a benchmark policy's waiting times, plus an abstract that doesn't match the tables.","tokens_in":39572,"tokens_out":3871,"would_cite":false,"duration_ms":41976,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90B06","90C11","90C15"],"pacs":[],"model":"deepseek-v4-flash","headline":"A sequential MILP charging policy that anticipates station queues and end-of-day energy needs raises electric ride-hailing profit by 7.65%-15.05% and service rate by 7-12.3 percentage points over four benchmark policies.","keywords":["ride-hailing","electric vehicles","charging scheduling","mixed integer linear programming","charging congestion","time-of-use energy prices","stochastic demand","fleet management"],"falsifier":"Re-run the five validation days after re-estimating the day-ahead plan's waiting-time and profit-rate inputs from the CongestionAware policy's own realized queues, iterating until the plan's predicted waiting times match the simulated ones; if the profit gap over the DynaThreshold benchmark falls below the reported 7.65%-15.05% or disappears, the advantage is an artifact of benchmark-calibrated parameters rather than of the scheduling logic.","tokens_in":38508,"feed_emoji":"⚡","tokens_out":6579,"duration_ms":61030,"temperature":0.7,"pith_summary":"The paper tries to show that an electric ride-hailing fleet can serve more customers and earn more profit without adding chargers, by scheduling charging around two things most earlier models ignore: queues at congested chargers and each vehicle's actual energy need to finish the day. It proposes a sequential mixed-integer linear program that first builds a day-ahead charging plan using expected waiting times and time-of-use energy prices, then reassigns vehicles to chargers online every minute. In simulations on a Manhattan-like grid with 100 EVs and 3,000 or 4,000 customers a day, the policy beats four benchmark charging rules on both profit and customer service rate, with profit gains of about 7.65% to 15.05% and service-rate gains of about 7 to 12.3 percentage points depending on scenario.","feed_headline":"Charging plan that predicts queues lifts fleet profit 15%","feed_subtitle":"Day-ahead schedules tuned to expected waiting times beat four benchmark charging policies in Manhattan-style tests.","key_machinery":"The load-bearing object is a sequential MILP with three interacting models: a day-ahead charging plan (P1) that assigns vehicles to half-hour epochs and target SoCs to minimize energy, access, and opportunity costs subject to charger capacity and a minimum charging time; a reactive pool-and-target rule (Eq. 29) that sets each vehicle's post-charge energy to just cover its remaining driving needs to end of day; and an online vehicle-to-charger assignment (P3) solved every minute that minimizes access, waiting, and charging time. The day-ahead plan is calibrated with expected waiting times and average profit rate estimated from simulating the Fastest benchmark policy on training days, which is what lets it steer charging away from predicted afternoon congestion peaks.","core_discovery":"On its own terms, the paper's central claim is that charging congestion, not charger scarcity, is the main drag on electric ride-hailing profitability, and that it can be managed by coordinating vehicles' charging times, durations, and charger choices with an eye on expected queueing. The proposed CongestionAware policy replaces the usual 'charge when the battery is low, charge to 80%' rule with a day-ahead plan that minimizes charging cost, access cost, and opportunity cost of time under charger capacity and minimum charging-duration constraints, and then adapts that plan online. In the reported experiments it achieves the highest profit of all five policies in every demand scenario and battery-size case, cuts total charging waiting time by roughly half to two-thirds, and concentrates end-of-day batteries near the reserve level instead of leaving the fleet half full.","pith_inferences":["The reported gains depend on calibrating the day-ahead plan with waiting times from a benchmark policy; a self-consistent calibration using the CongestionAware policy's own realized queues could either shrink or enlarge the advantage, and the paper does not provide that fixed-point check.","The partial-recharge rule effectively aims every vehicle's end-of-day battery at the reserve level; the same 'charge just enough to finish the shift' principle could apply to other shift-based fleets, from delivery vans to airport shuttles.","Because slow chargers are rarely used, the policy suggests that adding fast chargers, not slow ones, is the binding investment lever; a testable extension is to re-run the sensitivity analysis with fast-only expansions.","The day-ahead model's dependence on historical average waiting times could be replaced by a learned predictor that accounts for the policy's own feedback, connecting this approach to model-based RL without changing the online MILP core."],"forward_implications":["Operators can raise profit by 7.65%-15.05% and service rate by 7-12.3 percentage points without buying more chargers.","The advantage over benchmarks widens when charging infrastructure is scarce and when battery capacity is larger, so the policy is most valuable in exactly the conditions where naive charging rules break down.","Total charging waiting time drops to roughly a third to a half of benchmark levels, and fast chargers run near full occupancy during peak hours while slow chargers are mostly unused.","The method scales: online assignment solves in about 0.03 seconds even for 20,000 requests/day, while the day-ahead plan can be solved approximately within a one-hour time limit.","Replacing full 80% charges with energy-need-based partial charges is the main driver of the gain; time-of-use price awareness contributes only a small additional saving."],"supporting_citations":[{"why":"Supplies the two-stage MILP charging-management baseline that this paper extends with explicit congestion and ToU prices.","marker":"Ma (2021)"},{"why":"Provides the sequential MILP with charging station capacity that motivates the day-ahead planning model P1.","marker":"Jamshidi et al. (2021)"},{"why":"Establishes the long-horizon/short-horizon charging planning structure used here.","marker":"Zalesak and Samaranayake (2021)"},{"why":"Defines the Nearest benchmark charging policy.","marker":"Bischoff and Maciejewski (2014)"},{"why":"Defines the MinChgOpT benchmark based on minimum charging operational time.","marker":"Ma and Xie (2021)"},{"why":"Defines the DynaThreshold benchmark and dynamic SoC thresholds used for comparison.","marker":"Ahadi et al. (2023)"},{"why":"Models two-stage charging coordination with current queueing state, a direct predecessor of the online assignment P3.","marker":"Yang et al. (2019)"},{"why":"Represents the uncapacitated-charging RL alternative whose optimistic assumption this paper corrects.","marker":"Al-Kanj et al. (2020)"},{"why":"Supplies the real-world charging infrastructure context and fleet electrification targets behind the case study.","marker":"Taxi & Limousine Commission (2022)"}],"fun_headline_variants":["Queue-aware charging boosts ride-hail profit up to 19%","Predicting charger queues lifts e-fleet profit and service","Day-ahead charging plans cut queue waits, lift profit 3.9%","Congestion-aware charging beats benchmarks by 20% service rate","Electric ride-hail profit gains from charger queue prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The day-ahead plan trusts that waiting times and profit rates estimated by simulating a different (Fastest) charging policy on training days still describe the congestion the new policy will actually create; if those estimates are wrong, the schedules could be miscalibrated and the reported gains might not hold.","fun_headline_variants_meta":{"raw":{"variants":["Queue-aware charging boosts ride-hail profit up to 19%","Predicting charger queues lifts e-fleet profit and service","Day-ahead charging plans cut queue waits, lift profit 3.9%","Congestion-aware charging beats benchmarks by 20% service rate","Electric ride-hail profit gains from charger queue prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00018,"raw_usage":{"total_tokens":1281,"prompt_tokens":898,"completion_tokens":383,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":295}},"tokens_in":514,"tokens_out":383,"duration_ms":4349,"temperature":1.0,"reasoning_tokens":295,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:29:14.678994+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the five validation days after re-estimating the day-ahead plan's waiting-time and profit-rate inputs from the CongestionAware policy's own realized queues, iterating until the plan's predicted waiting times match the simulated ones; if the profit gap over the DynaThreshold benchmark falls below the reported 7.65%-15.05% or disappears, the advantage is an artifact of benchmark-calibrated parameters rather than of the scheduling logic.","supporting_citations":[],"review_version":1}